Train & Fine-Tune Models
9x Model Compression for Edge
INT8 quantization-aware training compressing MobileNetV2 from 23.5MB to 2.6MB with only 3.8% accuracy drop.
The problem
Models that perform well in the lab are too large to run on edge and mobile hardware without aggressive compression.
- The model that works great on your laptop won't fit on the target device at all.
- Naive post-training quantization tanks accuracy and nobody can say why.
- Every "make it smaller" pass risks breaking the exact behavior users rely on.
What NEO built
NEO ran INT8 quantization-aware training on MobileNetV2 using TensorFlow Lite, with a calibration dataset and augmentation to preserve accuracy through compression.
The result
9x model compression
Compressed the model 9x (23.5MB → 2.6MB) with only a 3.8% accuracy drop, ready for mobile and IoT deployment.
From the blog · 8 min
9x Model Compression with Quantization-Aware Training for Edge Deployment
How we compressed MobileNetV2 from 23.5MB to 2.6MB using INT8 quantization-aware training, achieving 9x size reduction with only a 3.8% accuracy drop, ready for deployment on Android, iOS, and Raspberry Pi.
Try this in your workspace
Paste this into NEO chat to kick off the same workflow on your own data.
Run quantization-aware training on my model to compress it for edge deployment, and tell me how much accuracy I'm actually trading away.
Paste it in · review the plan · get the diff
Get NEO
